Unsupervised AutoML and Dimensionality Reduction for Autonomous DDoS Attack Prediction

Machine learning models and feature selection are crucial for predicting Distributed Denial of Service (DDoS) attacks. Predicting attacks with high accuracy allows security teams to reduce attack damage. However, diversity in attacks and models limits predictions. Moreover, the dependence on labeled data and the utilization of unexpressive features restrict the performance of prediction models. This work proposes the AUTO-SEE technique to solve this problem. The technique engineers new features to reveal signals of attack preparation and selects the best features and the optimal machine learning model without using labeled data. This enables the technique to operate autonomously and predict different DDoS attack types, also increasing the protection against 0-day attacks. The results indicate that AUTO-SEE reduces error by up to 44.15%, reaching an accuracy between 72.41 and 100% in predicting DDoS attacks.